[torchlib] Fix index_put: handle None cases

[AyoubMDL]

This PR introduces support for None indices in the index_put function. If an index is None, it acts as a full slice (:).

index_put Logic:

1. Construct index grid that contains all the indices to be updated
2. Reshapes the update values to match the computed indices.

[AyoubMDL]

@microsoft-github-policy-service agree company="Brainchip"

[AyoubMDL]

@justinchuby PR is ready for review

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 72.24%. Comparing base (89dd454) to head (959dab4).
Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2061      +/-   ##
==========================================
- Coverage   72.35%   72.24%   -0.11%     
==========================================
  Files         216      217       +1     
  Lines       28890    29053     +163     
  Branches     3433     3449      +16     
==========================================
+ Hits        20904    20990      +86     
- Misses       6856     6934      +78     
+ Partials     1130     1129       -1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

Thank you! Could you look at the failing tests?

[AyoubMDL]

Force pushed as I changed a little bit the logic. Also added more test cases.

[justinchuby]

Could you take a look at the falling CI, and we can merge this PR? Thanks!

[AyoubMDL]

Could you take a look at the falling CI, and we can merge this PR? Thanks!

it was missing a case with multidimensional index (added a test for that case). Tell me if other cases are missing that you want to support.

[justinchuby]

For range, I think we need to Squeeze the tensor to remove its dimension.

[AyoubMDL]

@justinchuby I wasn’t able to make this work on symbolic tensors. Please feel free to close the draft or continue from here if helpful.

[justinchuby]

Thanks. Could you share the issues you have seen? I can take it from here

[justinchuby]

Will merge as-is and create follow up improvements

[AyoubMDL]

Thanks. Could you share the issues you have seen? I can take it from here

• I didn't know how to do the part where I construct the reshape_list using only onnx operations.
• I also tried to adapt the implementation from pytorch https://github.com/pytorch/pytorch/blob/6c3492b49118d9407db39fc5b134c0f96739b831/torch/onnx/symbolic_opset11.py#L206 but there logic doesn't work when I want to handle None indices.